Using Fuzzy Logics to Determine Optimal Oversampling...
Transcript of Using Fuzzy Logics to Determine Optimal Oversampling...
Using Fuzzy Logics to Determine Optimal Oversampling Factor for
Rasterizing RT Structures in DVH Computation
Csaba Pinter1, Tim Olding2, L. John Schreiner2, Gabor Fichtinger1
1Laboratory for Percutaneous Surgery, School of Computing, Queen’s University, Kingston, ON, Canada
2Department of Physics, Engineering Physics, and Astrophysics, Queen’s University, Kingston, ON, 5
Canada
Corresponding author: Csaba Pinter, [email protected]
Abstract
Purpose: Rasterizing three-dimensional surfaces into binary image volumes is a frequently performed 10
operation in radiation therapy (RT) workflows. For example, in both the clinic and research, dose-volume
histograms (DVH) are used to evaluate the quality of an RT treatment plan. To calculate a DVH, the 3D
surfaces (i.e. RT structures, usually targets and organs at risk) need to be rasterized into binary volumes.
The details of this step may significantly influence the output DVH, so special attention is needed when
setting up the rasterization parameters. Methods: An effective way of improving the quality of the 15
rasterized volume (i.e. increasing similarity between that and the original structure) is to apply
oversampling on the reference volume to simply increase the resolution of the output structure binary
volume. However, increasing the oversampling factor significantly raises the computational and storage
cost. This paper proposes a fuzzy-based automatic calculation of the oversampling factor so that higher
values are only used in cases where necessary and beneficial. A fuzzy inference system was introduced 20
that applies fuzzy rules to determine an optimal oversampling factor based on two measures: relative
structure size and structure complexity. Results: The proposed algorithm was used to automatically
calculate oversampling factor for chosen structures of three RT studies, two phantoms and one real
patient. The results show that the method is able to find the optimal oversampling factor in most cases,
and the calculated DVHs show good match to those calculated using manual overall oversampling of two. 25
Conclusion: A method has been presented to automatically determine rasterization oversampling factors
for RT structures. It ensures maintaining high DVH quality, enabling the planning system to find the
optimal treatment plan faster and more reliably.
Keywords: Radiation therapy, DVH, Rasterization, Fuzzy logics.
Introduction 30
A basic tool for evaluating treatment plans is the dose-volume histogram (DVH) that provides dose
information for target and organ-at-risk structures that is easily assessable by humans. An early step in
DVH calculation is converting the RT structure data that is stored in series of 2D planar contours into 3D
binary volumes, i.e. rasterization. These planar contours are often converted into 3D surfaces as a pre-
processing step so that rasterization can be performed more easily. Although the rasterization step itself 35
is then rather straightforward, there are certain decisions to be made that may significantly influence the
resulting DVH. One such decision is whether to include each border voxel, i.e. voxels of which only a part
falls inside the structure. This is an especially important factor in areas of high dose gradient, where
inclusion or exclusion of a single border voxel may influence the DVH in such a way that it compromises
the treatment plan optimization process, potentially resulting in a sub-optimal plan being selected by the 40
planning system.
Another decision to be made is the resolution of the grid that is used for rasterization. Often the
accompanying dose distribution volume is used as a grid reference, but its resolution is unsatisfactory in
case of smaller or more complex structures, where a significant proportion of the voxels are border voxels.
Using a too low resolution rasterization grid can introduce several types of artifacts in the DVH especially 45
in high dose gradient areas, such as a staircase-like effect, as well as consistent shift, as illustrated in Fig. 1.
As the grid resolution directly influences the proportion of border voxels, which we established above as
critical regarding the measurements, by increasing the grid resolution, the DVH inaccuracies can be
mitigated. In this situation, dividing the reference voxels into smaller equally sized voxels (i.e.
oversampling) seems to be a good idea, because then the borders of small or complex structures can be 50
approximated more accurately. The number of divisions along each axis is called the oversampling factor.
However, simply increasing the oversampling factor for each structure is not feasible, because it poses a
greater computational and storage demand, which may increase the processing time beyond the clinically
acceptable range. To mitigate this issue, determination of the oversampling factor for each structure is
necessary. Application of such per-structure rasterization grid resolution was also suggested by Drzymala 55
et al. [1] in their overview paper about DVHs. Although manual selection of these factors is a possibility,
it would not be feasible due to i) the added unnecessary complication to the planning process, ii) the
required additional staff training, and iii) its time requirements beyond clinical applicability.
Fig. 1: DVHs for the high resolution (solid) and low resolution (dashed) labelmaps showing the staircase effect of rasterization (left) and consistent shift (right).
We propose a method for automatic determination of the oversampling factor, so that the DVH 60
calculation process only uses high oversampling factors in case of the aforementioned problematic small
or complex structures, but uses the standard reference grid (factor of one) for the structures that are
relatively insensitive to the grid resolution, or even subsamples the volume for especially large structures
(factor of one half). The automated process aims to increase DVH quality in a clinically feasible way, while
leaving the operator planning workflow unchanged. 65
Methods
A few factors may influence the choice of the magnitude of oversampling, the two most significant of
which are relative structure size and structure complexity. The smaller a structure, the greater the ratio
of the border voxels among all structure voxels, which means greater uncertainty of volume measurement
and dose value inclusion (or exclusion). In these cases we can increase the accuracy by increasing the 70
oversampling factor. In the case of large structures the opposite is true, so using low oversampling values
can be used to reduce computation time, while not sacrificing accuracy. The other deciding factor is
structure complexity, as it also influences the ratio of border voxels. If a structure is simple, then most of
the included voxels fall inside the volume, while if it is complex, then the border voxels are more
emphasized. 75
Note that we have used terms such as “relatively small”, and “complex” which may be straightforward to
decide for a human, but less quantifiable. A natural choice of method for making a limited-choice decision
based on relatively subjective terms is the use of a fuzzy inference system, which is a mathematical system
that uses fuzzy set theory to analyze analog input values (features) in terms of logical variables to map
them to discrete output classes. 80
Fuzzy inference systems
As the following sections describe the details of the implemented fuzzy inference system, a brief
description of this concept is in order. Fuzzy sets [2] are special types of sets where the elements have
degrees of membership instead of the binary membership assessment in classical set theory. The
elements of a fuzzy set are assigned real values in the interval [0, 1], while elements of the traditional 85
“crisp sets” can only be assigned either 0 or 1. The fuzzy set theory allows representing more realistic
scenarios where the membership of an element in a set can be partial. Fig. 2 illustrates this concept
through defining fuzzy sets for age groups. Fuzzy logic defines fuzzy operators for the well-known Boolean
operators and, or, etc., as well as implication operators, which are the bases of the different fuzzy
inference systems, as described by Lee et al. [3]. These systems – also called as fuzzy control systems or 90
fuzzy rule systems – are designed to make decisions based on a number of fuzzy inputs. Usually crisp real-
world measurements are “fuzzified”, i.e. converted to membership values, which are the inputs (called
antecedents) to the inference system. Next, a pre-defined set of rules are applied on the antecedents to
determine the “consequents”, which are also membership values, but of the output fuzzy sets
representing the decisions. The fuzzy rules are most commonly intuitive if-then statements, which 95
naturally follow from real-world observations. The decision made by the system is a crisp value that is
gained by “defuzzifying” the consequents. This process consists of creating the weighted consequent
functions using the membership values, then combining the consequent functions into a single output
function, which is finally converted into a crisp value by a simple geometric operation, such as mean of
maxima or center of mass. The various fuzzy inference systems differ in the details of this process, such 100
as the implication function itself, the method of applying the weights to the consequent functions, or the
way of defuzzification. Detailed description of the applied inference system follows after specifying the
inputs and the output decisions.
Fig. 2. Illustration of fuzzy sets through age groups. The difference from crisp sets is that e.g. for the “young” set, the elements (i.e. people) are not divided by a “crisp” value (e.g. 26 years) to “completely young” and “not young at all”, but can have various degrees of youth. Furthermore, people can belong to multiple age groups with various degrees of memberships.
Structure representations 105
RT structures are stored in series of 2D planar contours, which can be processed in different ways so that
they properly enclose the contoured volume for rasterization (see Fig. 3). One method is to thicken the
contours along the axis orthogonal to the contour plane, thus creating “ribbons” (Fig. 3.B). Although a
ribbon model is not strictly a closed surface (Fig. 3.C), it can be used for rasterization under the right
conditions. These conditions may be reached by applying a rigid transformation, provided the contour 110
planes are parallel to each other. Another method for creating a contour is careful direct triangulation of
the contour points to create a closed surface. Such a method is described by Sunderland et al. [4]. A third
method is through pre-rasterization of the ribbon model and applying the marching cubes algorithm,
optionally followed by decimation and/or smoothing. This method is less appropriate in our case, as our
goal in the first place is to find a rasterization parameter, and also because pre-rasterizing a contour just 115
to calculate the final oversampling factor introduces major overhead, making it too slow for clinical use.
The above described contour and surface representations are reviewed in Fig. 3.
Fig. 3. Representations of a structure (brain stem). A: Original planar contours, B: Ribbon models, C: Closed surface triangulated directly from planar contours, D: Close surface via pre-rasterization
Relative structure size
The main measure based on which the optimal oversampling factor is calculated is relative structure size, 120
which we chose to define as the ratio of the structure volume and the volume of the grid reference image,
i.e. the volume enclosed by the bounding box of the reference CT. It is satisfactory to consider the
structure size compared to the reference volume, as the interest is in the relative size of an image voxel
compared to the structure.
Four fuzzy sets describing the relative structure size (RSS) can be defined (see Fig. 4): 125
1. Large – body, lung, group of medium-sized organs, large isodose volume, etc.
2. Medium – bladder, heart, brain, group of small organs, etc.
3. Small – brain stem, eyeball, spinal canal, etc.
4. Very small – optic lens, optic chiasm, optic nerve, hotspot isodose volume, etc.
These volume ratio values range from the magnitude of 10-6 to almost one. As this range cannot be 130
partitioned well in a linear way, a different scale needs to be used. Since the relative structure size is quite
well described by the power of ten of this fraction, we decided to use a base ten logarithmic scale. The
final measurement considered for the fuzzy rules is
𝑅𝑆𝑆 = −𝑙𝑜𝑔10(𝑣𝑜𝑙𝑢𝑚𝑒𝐹𝑟𝑎𝑐𝑡𝑖𝑜𝑛) (1)
which means that the relative structure size will go from zero to five as the size increases.
Structure complexity (compactness) 135
The second input measure for the fuzzy inference system is structure complexity, which is a term that is
quite hard to define quantitatively. According to Li [5], “the meaning of this quantity should be very close
to certain measures of difficulty concerning the object or the system in question: the difficulty in
constructing an object, the difficulty in describing a system, the difficulty in reaching a goal, the difficulty
in performing a task, and so on”. 140
The topic of quantifying complexity of spatial objects has been studied by several groups. Brinkhoff et al.
[6] proposed a 2D contour complexity calculation method mainly for geographical formations. It could be
useful for this work if the datasets were always expected to have 2D contours as input, which is not the
case, as the resulting algorithm is intended to support shapes from other sources as well, such as DICOM®
SEG (Digital Imaging Communication in Medicine1) or non-DICOM data. O’Rourke et al. [7] developed a 145
method for finding minimal enclosing boxes for shapes, which could potentially be useful to apply for
complexity measure computation, but aside from the problem mentioned with the Brinkhoff paper, it
would also be too computationally complex. There are other results, such as the method of Rigau et al.
[8], which uses mutual information to measure object complexity, but it is also too computationally
demanding for a task where we actually strive to save computation time by calculating these measures. 150
Another possibility for quantifying complexity would be to compute a similarity metric (Dice similarity [9]
or Hausdorff distance [10]) of the structure and the sphere centered at the structure’s center of mass,
with a volume equal to that of the structure. However, these similarity measures require rasterized inputs,
so using them would be a major overhead.
The measure of complexity we chose for our algorithm is of Alyassin et al. [11], which is specifically 155
designed for medical data, and is implemented in the VTK [12] toolkit in the filter vtkMassProperties. This
algorithm calculates the normalized shape index (NSI), which characterizes the deviation of the shape of
an object from a sphere. A sphere's NSI is one by definition and the NSI is always greater than or equal to
one for complex structures.
In experiments conducted on typical anatomical structure sets from the SlicerRT data repository [13] 160
(which includes the datasets detailed below) the NSI value fell between 1 and 2, so we chose
𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝑖𝑡𝑦 = 𝑁𝑆𝐼 − 1 (2)
as the input complexity measure for the fuzzy rule system. For example, the eyeballs have complexity
value of zero, and branching vessels are assigned a value of 0.6. One more consideration is important
when working with NSI. The algorithm assumes that its input is a closed surface, which is not true in case
1 DICOM is the registered trademark of the National Electrical Manufacturers Association (NEMA) for its standards publications relating to digital communications of medical information.
of ribbon models, as only the “vertical” components of the structures are triangulated. Our experiments, 165
however, show that the NSI of the ribbon model of a structure is more or less consistently 85% (falls
between 84% and 91%, and the higher values are more rare) of that of the NSI for the actual closed surface
of the same structure, which suggests that NSI measurement are reliable for ribbon models, too.
Two fuzzy sets were defined to describe the complexity of a structure: Low and High (see Fig. 4). The fuzzy
set parameters were determined based on metrics calculated on a set of structures that were subjectively 170
classified.
Output oversampling factor
Based on our experience in defining the used rasterization grid by oversampling the dose volume, only a
few sensible oversampling factors exist. Too low oversampling factors make an overly crude rasterization
grid resulting in unacceptable data loss even for the largest and simplest structures, while the 175
computational gain is less significant with every decrease of the factor. On the other hand if a too large
factor value is chosen, the rasterization step will take too long to be feasible, while the geometric gain will
become negligible. Thus four possible values were chosen for oversampling factors: ½, 1, 2, and 4.
The fuzzy sets of the input measures and the output oversampling factor are reviewed in Fig. 4. The input
membership functions represent the degree of membership of each structure’s quantified property in the 180
defined fuzzy sets.
Fig. 4: Fuzzy memberships for the input measures and for the output oversampling factor.
Fuzzy inference system
A Mamdani-style fuzzy inference system [3] has been implemented defining the following rules:
1. If RSS is Very small, then Oversampling is Very high 185
2. If RSS is Small and Complexity is High then Oversampling is High
3. If RSS is Medium and Complexity is High then Oversampling is High
4. If RSS is Small and Complexity is Low then Oversampling is Normal
5. If RSS is Medium and Complexity is Low then Oversampling is Normal
6. If RSS is Large, then Oversampling is Low 190
First the input measurements are fuzzified using the membership functions, i.e. the measured exact value
is converted into membership value for each set. Next, the rules are applied on them, yielding the output
(consequent) functions that are clipped to the membership values, and finally their center of mass is
computed, which is the output “crisp” value. This crisp resulting oversampling factor power is rounded,
and the actual oversampling factor is calculated as follows: 195
𝑜𝑣𝑒𝑟𝑠𝑎𝑚𝑝𝑙𝑖𝑛𝑔𝐹𝑎𝑐𝑡𝑜𝑟 = 2𝑟𝑜𝑢𝑛𝑑(𝑐𝑒𝑛𝑡𝑒𝑟𝑂𝑓𝑀𝑎𝑠𝑠) (3)
Seeing how the inference system works in practice helps in understanding. We chose to illustrate the
process through the example of the Orbit Lt structure as seen in Table 1. The input crisp measurements
are 3.59 for RSS, and 0 for complexity (as the eyeball is spherical). After fuzzification, the membership
values for the RSS sets are 0, 0, 0.32, and 0.68 for the sets Large, Medium, Small, and Very small,
respectively, and 1 and 0 for the complexity sets Low and High, respectively. Based on the six rules, the 200
consequent values that are used for clipping the output functions are 0.68 for the first rule (because this
is “how much” RSS is Very small), 0 for the second rule (because RSS is Small to the degree of 0.32, but
complexity is High to the degree of 0, and the and operator takes the minimum of the operands in the
Mamdani system), 0 for the third rule, 0.32 for the fourth rule, and 0 for the last two rules. After applying
the consequent values to the output sets based on the rules, the consequent functions are formed as seen 205
in Fig. 5. The center of mass of the consequent functions is calculated to be 1.08, and so Eq. 3 yields the
value of 2, which is the defuzzified output.
Fig. 5: Consequent functions for the Orbit Lt structure used in the example. The first rule resulted in clipping the Very high output function to 0.68, the fourth rule clipped the Normal function to 0.32, and the rest of the rules resulted in clipping to 0. The center of mass of these consequent functions is 1.08.
Implementation
The fuzzy inference system was implemented as part of the SlicerRT toolkit [13]. It is a general radiation 210
therapy research toolkit based on 3D Slicer [14], an open-source medical image analysis and visualization
platform. The fuzzy oversampling calculation algorithm was implemented as one C++ class
(vtkCalculateOversamplingFactor) that integrates in the general SlicerRT contour handling mechanism and
the DVH module. All information regarding the development can be found in the project’s issue tracking
system: https://www.assembla.com/spaces/slicerrt/tickets/189. 215
Results
Three RT studies were used to test the method: a treatment plan study for the RANDO® prostate phantom
(referred to as “PROS”), a treatment plan study for the RANDO® ear-nose-throat phantom (“ENT”), and a
treatment plan for a real patient that contains elaborate contours (“Branching”). We also tested the
algorithm on a non-RT MR study that contained a structure with tilted planes (“Tilted”) to verify our 220
method on this special acquisition type. A “quantitative” workspace with one chart view for a DVH
computation for the ENT dataset can be seen in Fig. 6. The results are shown in Table 1.
Fig. 6. A SlicerRT “quantitative” workspace with one chart view showing the comparison of the DVH for ENT structures with constant manual oversampling value of 2 (solid line) and automatic oversampling (dashed line). The difference between the default and the calculated oversampling is visible for the small structures even fully zoomed out.
225
Structure Approximate
shape
Relative structure
size
Subjective size
Complexity measure
Subjective complexity
Sampling model
G.T. R.O. CvR.O. CfC.O.
PROS
Bladder spherical 2.17 Med 0.09 Low 1 1 1 1
Rectum cylindrical-vertical 2.78 Med 0.35 Med 2 1 * 2 1 *
Body cylindrical 0.26 Large 0.13 Low 0.5 0.5 0.5 0.5 Femoral Head Lt spherical 2.57 Med 0.11 Low 1 1 1 1
ENT
Brain spherical 1.50 Med/Large 0.15 Low 1 1 1 1 Brain Stem
cylindrical-vert. (banana) 3.08 Small 0.30 Med 2 1 * 1 * 1 *
Lens Lt hemisphere 5.17 Very small 0.02 Low 4 4 4 4 Optic Chiasm
"bone-shaped" 4.20 Very small 0.05 Med 4 4 4 2 *
Optic Nerve Lt
cylindrical-horizontal 3.89 Very small -0.04 Low/Med 4 4 4 1 -
Orbit Lt spherical 3.56 VeryS/Small 0.09 Low 2 2 2 2
PTV1 spherical 2.42 Med 0.08 Low 1 1 1 1
Branching
Body torso 0.35 Large 0.38 Med 0.5 0.5 0.5 0.5
Kidney Rt kidney 2.66 Med 0.23 Low/Med 2 2 1 * 1 *
Vessels branching cylindrical (Y) 2.61 Med 0.99 High 2 2 2 2
Spinal Canal
cylindrical-vertical-long 3.27 Small/Med 0.52 Med/High 2 2 2 2
Liver liver 1.97 Med/Large 0.47 Med/High 1 2 * 1 1
Tilted
Target cylindrical-wide 2.19 Med 0.31 Low 2 1 * 1 * 1 *
Table 1. Test structures, their computed and subjectively determined measures, and their computed oversampling factors. The relative structure size and complexity measure show the averaged values for the three representation types. Sampling model abbreviations (see Fig. 3 for reference): G.T. – Ground truth (oversampling factor), R.O. – Ribbon oversampling (automatically calculated from the ribbon models), CvR.O. – Closed surface via rasterization oversampling (using the closed surface created from pre-rasterized volumes with oversampling of two), CfC.O. – Closed surface from contour oversampling (using the closed surface directly triangulated from the planar contours). Color-codes: green (no mark) – perfect match, yellow (*) – acceptable, red (-) – unacceptable.
Ground truth oversampling factors for each test structure were determined as consensus by several
clinicians in the cancer program at the Kingston General Hospital. Each of the participating clinicians can
be considered an expert in RT treatment planning, and the decision was made based on their personal 230
experience. The decision regarding the optimal oversampling factor involved considering the following
properties of the individual structures: typical size and complexity, variance in size and shape within the
patient population, and proximity of the structure to high dose gradients. The calculated oversampling
factors using the three different surface representations were compared to the ground truth. The results
show good matching overall, especially considering the ribbon models and the closed surfaces via pre-235
rasterized volumes. The one failure happened with the closed surfaces directly generated from the planar
contours.
The DVHs computed using automatic oversampling factors and a constant manual factor of two were also
compared to results of third party applications via SlicerRT’s DVH comparison module. The comparison
algorithm follows the gamma-like DVH comparison method described by Ebert et al. [15]. We used the 240
results from two applications, one being CERR [16], a popular Matlab-based radiation therapy research
tool, and Eclipse® (Varian Medical, Palo Alto, CA, USA), a commercial RT treatment planning system that
is used in many hospitals and clinics all over the world. According to the CERR documentation no
oversampling is performed, which means the software uses a factor of 1. Unfortunately we were not able
to access such documentation for Eclipse®, and the software gives no information about the DVH 245
calculation details. However, its DVH calculation has been validated [17], and was found to satisfy the
requirements set by the authors, so can be considered a suitable baseline for this study. The comparison
results (see Table 2) show that the new method utilizing automatic oversampling factor calculation
provides a better overall match to the DVHs calculated by the third-party applications than the use of a
fixed oversampling factor of two. The cases with lower match percentages are all structures in high dose 250
gradients (such as rectum and brain stem), where sampling makes a larger difference. These low match
values may be accounted for the very strict tolerances that were used to be able to detect even the
smallest deviations (1%/1mm instead of the typical gamma tolerances 3%/3mm). The DVH computation
time including rasterization and the calculation of oversampling factor (if applicable) was, as expected,
less in case of lower overall oversampling factors than the default of two (this was the case with PROS, 255
average of 1.6 s for manual and 0.4 s for automatic), and more for higher overall oversampling (with ENT,
average of 5.2 s for manual and 10.7 s for automatic).
Structure Eclipse® w/ Auto
Eclipse® w/ Manual
CERR w/ Auto
CERR w/ Manual
PROS Overall 95.74 94.34 99.93 97.54
BODY 100 100 100 100
Bladder 100 100 100 100
Femoral Head Lt 99.9 99.9 99.58 100
Femoral Head Rt 98.14 99.9 100 100
PTV 99.61 99.8 100 100
Rectum 76.76 66.41 100 85.23
ENT Overall 93.63 91.4 96.39 94.25
BODY 100 100 100 100
BRAIN 100 100 100 100
BRSTEM 74.81 43.3 93.63 65.29
CTV 100 100 100 100
GTV 99.04 99.04 99.04 99.04
Lens Lt 98.56 98.56 98.73 98.73
Lens Rt 99.52 99.52 99.36 99.36
Optic Chiasm 82.18 81.03 85.67 81.85
Optic Nerve Rt 86.49 86.59 85.99 85.99
Optic Nerve Lt 85.73 85.73 85.67 85.67
Orbit Lt 97.8 97.8 99.36 99.96
Orbit Rt 98.66 98.66 98.73 98.73
PTV 100 99.14 100 100
optBRAIN 100 100 100 100
optOptic 81.61 81.61 99.68 99.68
Table 2. DVH comparison for SlicerRT manual oversampling of 2 (columns with
Manual) and automatic oversampling (columns with Auto), to CERR and Eclipse®
The computation time of the oversampling calculation algorithm itself was also assessed, by measuring
the time needed to perform the two input measurements, and the time needed to run the fuzzy inference 260
system. 178 measurements were made with all the three types of input surfaces and all the test
structures. The summarized average computation times can be seen in Table 3.
Total time (s) Measures (s) Fuzzy rules (s)
Mean 0.0165 0.0059 0.0106 Std.Dev. 0.0161 0.0145 0.0041 Min 0.0060 0.0000 0.0050 Max 0.1120 0.0990 0.0370
Table 3. Overall averaged time measurements for calculating the automatic oversampling factor on each structure listed in Table 1.
3D Slicer can be downloaded from http://download.slicer.org for Windows, Mac, and Linux 64-bit
operating systems. SlicerRT can then be installed from the Extension Manager in 3D Slicer. SlicerRT 265
versions 0.15.2 and above contain the described method. Information about usage can be found at
http://slicerrt.org. The software comes with BSD-style license, so it is freely usable, modifiable and
distributable.
Discussion
The ribbon model and closed surface via pre-rasterization show the most consistent results from the three 270
examined methods. The deviation concerning the closed surface from planar contours representations
that resulted in the unacceptable automatic factor for the optic nerve may be caused by the current
method of “end-capping”, which simply extends the closed surface by half the slice thickness at the top
and bottom contours to fully encapsulate the volume intended when delineating the structure on the
anatomical images. Implementing a more elaborate method of closing the extremes may result in better 275
match, especially in the case of structures such as the optic nerve structure in the ENT dataset, which
spans only three slices. Despite yielding good results, closed surface via pre-rasterization is not a viable
option as the goal itself is rasterization, and rasterizing just for the sake of determining the oversampling
factor introduces a major overhead. Thus although the ribbon models cannot be used in a completely
reliable way with the complexity measurement algorithm, the experienced consistency seems 280
satisfactory.
A possible improvement could be using other complexity measures. The current one only calculates a
global complexity measure, but we may also be interested in local complexity, so that locally highly
complex regions (“zig-zags”) are not oversimplified during rasterization.
It would be also interesting to see how odd oversampling values perform versus the currently used even 285
numbers. As the paper by Nelms et al. [18] suggests that using odd values ensures that “no slabs of ‘new’
voxels (that will need to be interpolated) are centered exactly between the original dose plane locations”,
this consideration might further improve the accuracy of the rasterized structure volumes.
Also, according to the comparison data in Table 2, the SlicerRT algorithm consistently shows better match
to CERR than to Eclipse® in case of elongated, string-like structures such as brain stem and rectum. It 290
would be worth investigating why this happens by examining more such structures, and also by
performing comparison using not only clinically sensible oversampling factors, but the highest achievable
resolution, thus creating a “gold standard”.
Conclusion
Accurate DVH calculation is essential for dose evaluation in both clinical and research settings. It is even 295
more so if inverse treatment planning is performed, because small or complex structures being involved
in the dose-volume objectives may skew the optimization process due to rasterization inaccuracies,
especially in areas of high dose gradient, potentially resulting in sub-optimal treatment plans. A fuzzy-
based method was implemented for automatically determining oversampling factor for each processed
structure to facilitate highly accurate DVH calculation while keeping computational cost reasonable. The 300
results show that the proposed algorithm reliably identifies the small and complex structures on various
input types and on a wide range of clinical structures including the extreme cases. It facilitates more
dependable dose-volume based optimization methods, while yielding a good balance between metric
accuracy and calculation time.
Disclosure of Conflicts of Interest 305
The authors have no relevant conflicts of interest to disclose.
References
1. R. E. Drzymala et al. "Dose-volume histograms". International Journal of Radiation Oncology* Biology*
Physics, 21(1), 71-78 (1991).
2. L.A. Zadeh, “Fuzzy sets”. Information and control, 8(3), 338-353 (1965). 310
3. C.C. Lee, "Fuzzy logic in control systems: fuzzy logic controller. II." Systems, Man and Cybernetics, IEEE
Transactions on 20.2 419-435 (1990).
4. K. Sunderland et al., “Reconstruction of surfaces from planar contours through contour interpolation”.
SPIE Newsroom. In press (2015).
5. W. Li. “On the relationship between complexity and entropy for Markov chains and regular 315
languages”. Complex Systems, 5(4), pp381–399 (1991).
6. T. Brinkhoff et al. "Measuring the Complexity of Polygonal Objects." ACM-GIS pp109-117 (1995)
7. J. O'Rourke et al. "Finding minimal enclosing boxes." International journal of computer & information
sciences 14.3, 183-199 (1985).
8. J. Rigau et al. "Shape complexity based on mutual information." Shape Modeling and Applications, 320
2005 International Conference. IEEE (2005).
9. L.R. Dice, "Measures of the amount of ecologic association between species." Ecology 26.3, 297-302
(1945).
10. D.P. Huttenlocher et al. "Comparing images using the Hausdorff distance." Pattern Analysis and
Machine Intelligence, IEEE Transactions on 15.9, 850-863 (1993). 325
11. A. M. Alyassin et al. "Evaluation of new algorithms for the interactive measurement of surface area
and volume." Medical physics 21.6 741-752 (1994).
12. W. Schroeder et al. “Visualization Toolkit: An Object-Oriented Approach to 3D Graphics”, 4th ed.
Kitware, Clifton Park, NY (2006).
13. C. Pinter et al. "SlicerRT: Radiation therapy research toolkit for 3D Slicer." Medical physics 39.10 6332-330
6338 (2012).
14. S. Pieper S. et al. “3D SLICER”. Proceedings of the 1st IEEE International Symposium on Biomedical
Imaging: From Nano to Macro; 632-635 (2004).
15. M. A. Ebert et al. “Comparison of DVH data from multiple radiotherapy treatment planning systems.”
Physics in medicine and biology 55.11, N337 (2010). 335
16. J.O. Deasy et al. "CERR: a computational environment for radiotherapy research." Medical physics
30.5: 979-985 (2003)
17. M.S. Gossman et al. "Dose-volume histogram quality assurance for linac-based treatment planning
systems." Journal of medical physics/Association of Medical Physicists of India 35.4, 197 (2010)
18. B. Nelms et al. "Methods, software and datasets to verify DVH calculations against analytical values: 340
Twenty years late (r)." Medical physics 42.8, 4435-4448 (2015).